Research

Gabriel Kasmi



Reliable and Scalable Deep Learning for the Energy Transition

My research explores how we can better understand, trust, and deploy deep learning systems to support the energy transition. I work at the intersection of explainable AI, geospatial data, and renewable energy systems, combining theoretical insights with applied work on the causes and consequences of rooftop photovoltaic (PV) development.

This work bridges algorithm design, large-scale mapping, and applied energy analytics, and is organized along three main threads. I highlight future research directions for each thread. If one of these directions is of interest to you, feel free to contact me!

Remote Sensing of Rooftop Photovoltaic Systems

A large share of PV capacity is distributed across rooftops, but these systems are often poorly mapped and monitored. In France, this stems from fragmented information spread across stakeholders, each with their own priorities and data formats.

The core of my thesis focused on mapping rooftop PV systems and building a nationwide registry. I developed DeepPVMapper, an algorithm for detecting and characterizing rooftop PV systems at scale, and released an open-source library, PyPVRoof, for further development. Using these tools, we mapped over 500,000 systems in France and validated the results against existing datasets—showing their potential to bridge critical information gaps as PV deployment accelerates.

Today, my focus is on making these mapping methods more actionable, reliable, and computationally efficient. One emerging direction is designing an optimal pipeline that balances accuracy, computational cost, and data availability. I’m also exploring mapping in new contexts, particularly in developing countries where PV is often adopted informally.

Future research directions:

Methods: Reliability and Interpretability of Machine Learning Models

The use of deep learning for computer vision is now commonplace. The real challenge lies not in implementation, but in deploying these models at scale and trusting their predictions.

My PhD was designed as a cookbook to improve the reliability of neural networks applied to the remote sensing of PV systems. See this paper for a summary of the approach and how interpretability and data augmentation can be used to improve the reliability of deep learning models.

I notably introduced a new explainability method, the Wavelet Scale Attribution Method (WCAM, XAI in Action workshop @ NeurIPS 2023), which reveals whether models rely on shapes, textures, or other structural features—offering richer insights than standard pixel-based saliency maps.

We later expanded this into the Wavelet Attribution Method (WAM, ICML 2025), which unifies feature attribution across different modalities (images, audio, etc.). Looking ahead, I’m exploring how to select meaningful feature attribution domains (with wavelets as one example) and whether we can combine concept-based explanations (powerful but abstract) with the practicality of feature attribution.

Future research directions:

Applications: PV Power Estimation and Socio-Economic Insights

Beyond methodological work, I apply these tools to tackle real-world energy challenges. The ultimate goal of rooftop PV mapping is to improve observability, enabling grid operators to produce precise estimates of rooftop PV generation.

The final chapter of my thesis introduced a simplified method for PV production estimation (see the paper here). We demonstrated that this method slightly outperforms the one currently used by the French transmission system operator (TSO) and, importantly, better accounts for self-consumption practices.

I have also co-supervised work leveraging DeepPVMapper data to analyze socio-economic patterns in PV adoption, from local to national scales.

Future research directions:

Selected Publications

Here are some of my key publications. For a complete list, see the Full Publication Record below.

Full Publication Record

Publications in peer-reviewed journals

International conferences proceedings (peer reviewed)

Workshops (peer reviewed)

Oral presentations

Position papers, working papers

Preprints

Posters

Miscellaneous works

This work in process was presented during the PhD Forum at ECML-PKDD 2022. It it a snapshot of our early attempts to use Fourier theory to explain the (lack of) robustness of a CNN classifier. This work later led to the WCAM. The manuscript is accessible here and the slides of the presentation here.

Reviewer

Journals

Conferences

Workshops